Multi-source heterogeneous data fusion group construction enterprise digital and intelligent integrated collaborative system

Abstract

The application discloses a multi-source heterogeneous data fusion group construction enterprise digital and intelligent integrated collaborative system, comprising a data acquisition layer, a data processing layer, a model calculation layer, an application service layer and a display interaction layer. The data acquisition layer acquires construction site video, Internet of Things sensor, personnel positioning, equipment state and environmental monitoring data; the data processing layer performs time-space alignment, cleaning, fusion and feature extraction; the model calculation layer performs safety risk time-space correlation analysis and probability prediction based on a space-time graph convolution network (ST-GCN); the application service layer triggers multi-level early warning, generates an emergency disposal scheme and intelligently distributes tasks; and the display interaction layer provides three-dimensional visualization and interaction functions based on digital twin technology. Through multi-source data fusion and time-space correlation analysis, the application realizes active prediction and fine control of construction safety risks, and improves emergency response efficiency and the intelligent level of safety management.

Description

Technical Field

[0001] This invention relates to the field of technology, specifically to a digital and intelligent integrated collaborative system for group construction enterprises that integrates multi-source heterogeneous data. Background Technology

[0002] With the rapid development of the construction industry and the increasing complexity of engineering projects, construction safety management is facing increasingly severe challenges. Modern construction activities involve a large number of personnel movements, heavy equipment operation, cross-operation of multiple trades, and complex and ever-changing working environments. These factors make the safety risks in the construction process diversified, dynamic, and interconnected.

[0003] On the one hand, existing construction safety monitoring systems primarily employ single-threshold alarm mechanisms. For example, in tower crane monitoring systems, alarm thresholds are typically set for single parameters such as wind speed and load; an alarm is triggered when the monitored data exceeds the preset threshold. However, safety risks in actual construction scenarios are often the result of the combined effects of multiple factors, and relying solely on a single indicator's threshold is insufficient to accurately identify complex safety hazards. On the other hand, currently widely used safety management systems suffer from significant deficiencies in terms of accuracy across spatial and temporal dimensions. Most systems can only provide a description of the overall safety status of the project or a large area, failing to achieve refined risk localization for specific work areas, equipment perimeters, or personnel activity areas.

[0004] Regarding the transmission and response mechanisms for early warning information, existing technologies primarily rely on manual transmission and a multi-level reporting process. When on-site monitoring equipment detects an anomaly, on-site safety personnel typically need to report to superiors via telephone, walkie-talkie, or other means before decisions are made and action instructions are issued. This multi-level information transmission chain is not only time-consuming but also prone to information distortion or delays during transmission, severely impacting the timeliness of emergency response. Furthermore, traditional safety management systems mainly rely on real-time monitoring data for post-event responses, lacking the ability to predict safety risks. Existing technologies also have significant shortcomings in the accumulation and transmission of safety management knowledge. Safety incidents and emergency response experiences during construction are often scattered across various documents and individual personnel's working memories, lacking systematic organization and structured storage.

[0005] Therefore, this solution proposes a digital and intelligent integrated collaborative system for group construction enterprises that integrates multi-source heterogeneous data to solve the above problems. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a digital and intelligent integrated collaborative system for group construction enterprises that integrates multi-source heterogeneous data.

[0007] To achieve the aforementioned objective, the technical solution of this invention is implemented as follows: a multi-source heterogeneous data fusion integrated digital and intelligent collaborative system for group construction enterprises, comprising a data acquisition layer, a data processing layer, a model calculation layer, an application service layer, and a display and interaction layer. The data acquisition layer is used to acquire video surveillance data, IoT sensor data, personnel positioning data, equipment status data, and environmental monitoring data from the construction site; the data processing layer is used to perform spatiotemporal alignment, cleaning, fusion, and feature extraction on the multi-source heterogeneous data; the model calculation layer includes a risk prediction model based on a spatiotemporal graph convolutional network (ST-GCN) for spatiotemporal correlation analysis and probability prediction of construction safety risks; the application service layer is used to trigger multi-level early warnings, generate emergency response plans, and intelligently distribute tasks based on the prediction results; the display and interaction layer provides a three-dimensional visualization interface and human-computer interaction functions based on digital twin technology.

[0008] Preferably, the spatiotemporal graph convolutional network adopts a dynamic graph update mechanism: the adjacency relationship of the spatiotemporal graph is dynamically updated every preset time according to the personnel location and equipment status at the construction site; the spatiotemporal graph nodes correspond to the grid units divided at the construction site, and the edges represent spatial adjacency relationships or functional association relationships; the network structure includes alternating stacked graph convolutional layers and temporal convolutional layers.

[0009] Preferably, the data processing layer includes a data robustness enhancement module, which calculates a quality score for each data source in real time (the scoring dimensions include data freshness, completeness, stability, and physical rationality), and dynamically adjusts the data fusion weights based on the score; for short-term missing data (continuously missing <10 sampling points), a spatiotemporal K-nearest neighbor imputation method is used; for long-term missing data (continuously missing ≥10 sampling points) or multi-dimensional joint missing data, data completion based on conditional generative adversarial networks (cGAN) is used; outliers are detected by cascading isolation forest and 3σ criterion, and median filtering or linear interpolation correction is performed according to the data type; a local cache is configured on the edge server, and the lightweight risk prediction model runs independently on the edge server when cloud communication is interrupted.

[0010] Preferably, the display and interaction layer constructs a digital twin model based on BIM, GIS and on-site 3D scanning data, and uses instantiation technology and level of detail (LOD) optimization to achieve model lightweighting, maps monitoring data to model objects in real time, and visualizes the distribution of safety risks through color changes, heat maps, animations and other methods; it supports four-dimensional spatiotemporal modeling and spatiotemporal backtracking functions.

[0011] Preferably, the application service layer adopts a multi-level early warning mechanism: based on the probability of risk, the degree of impact, and the urgency, the security status is divided into four levels: normal, attention, early warning, and emergency, with each level corresponding to differentiated response measures. After an early warning is triggered, the system provides intelligent decision support based on a knowledge base. The knowledge base uses a hybrid storage of graph database (Neo4j) and vector database (Milvus), containing 2150 deterministic rules and more than 500 typical cases, supporting a hybrid reasoning mechanism that combines rule reasoning and case reasoning. The knowledge base adopts a continuous learning mechanism, automatically adding new cases after each security incident is handled, extracting new rules through association rule mining weekly, and cleaning up redundant outdated cases.

[0012] The beneficial effects of this invention are reflected in: By integrating multi-source data and spatiotemporal correlation analysis, and comprehensively considering the synergistic changes of multiple related factors to assess risks, as well as shortening the response time of intelligent task distribution and collaborative mechanisms, this invention's deep learning-based risk prediction model can identify potential risk trends in advance, enabling early warning and proactive prevention. The digital twin technology of this invention supports multi-scale refined management from the overall to the local level, and can accurately locate specific components, equipment, and work areas. The full-process digital recording and structured knowledge base promote cross-project sharing and continuous optimization of safety management experience. Attached Figure Description

[0013] In the attached diagram: Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a structural diagram of a risk prediction model based on a spatiotemporal graph convolutional network. Figure 3 Flowchart for multi-source data fusion and situational awareness; Figure 4 A schematic diagram illustrating the mapping between a digital twin model and real-time data; Figure 5 A flowchart for multi-level early warning and intelligent decision support; Figure 6 A schematic diagram of a coordinated emergency response and closed-loop management mechanism. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of them. Unless otherwise specified, the embodiments and features described in this application can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0015] It should be noted that if the embodiments of the invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0016] Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the invention.

[0017] Example 1: A Construction Safety Risk Prediction Method Based on Spatiotemporal Graph Convolutional Networks (Fall Risk of High-Rise Buildings) See appendix Figures 1 to 6 This invention provides a digital and intelligent integrated collaborative system for group construction enterprises that integrates multi-source heterogeneous data. Among other things, Figure 1 A schematic diagram of the overall architecture of the system is shown. Figure 2 This is a structural diagram of a risk prediction model based on a spatiotemporal graph convolutional network (ST-GCN). This embodiment uses fall risk prediction during the construction of high-rise buildings as an example to illustrate the specific implementation of the ST-GCN-based safety risk prediction method.

[0018] In terms of data acquisition, this embodiment adopts the following technical configuration: (1) High-definition cameras are deployed on the building facade and work area, with a video stream resolution of 1920×1080 and a frame rate of 25 frames / second, for real-time monitoring of personnel behavior and object status; (2) Each construction worker wears a UWB (ultra-wideband) positioning tag with a positioning accuracy of 0.5 meters and a positioning data update frequency of 1Hz; (3) Wind speed sensors, tilt sensors and vibration sensors are deployed at key locations with a sampling frequency of 10Hz; (4) An edge computing server is deployed, configured with an NVIDIA Tesla T4 GPU, for local data preprocessing and model inference. In terms of model configuration, the spatiotemporal graph convolutional network adopts a 6-layer graph convolutional structure and a 2-layer temporal convolutional structure, with an input time window length of 60 minutes and a time step of 1 minute. The number of nodes in the graph is dynamically determined according to the number of hazardous sources divided at the construction site, usually 50 to 200 nodes. The model training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a training period of 100 epochs.

[0019] Step 1: Construct the spatiotemporal diagram structure of the construction site (see...) Figure 2(See the diagram for the graph structure). First, based on the building's spatial structure and functional zoning, the construction site is divided into multiple grid units, each serving as a node in the spatiotemporal graph. Edges between nodes represent spatial adjacency or functional relationships, such as adjacent grids on the same floor or corresponding vertical grids on different floors. The graph structure is represented by an adjacency matrix A, where... This indicates that there is a connection between node i and node j. This indicates no connection. This embodiment introduces a dynamic graph update mechanism: the system dynamically updates the adjacency relationships of the graph every 5 minutes based on personnel locations and equipment status, enabling the graph structure to reflect the actual situation on site in real time.

[0020] Step 2: Multi-source data feature extraction and spatiotemporal alignment. Human behavior features are extracted from video surveillance data, such as using a posture estimation model to identify whether workers at heights are correctly wearing safety belts; personnel location features and movement trajectory features are extracted from personnel positioning data; environmental features such as wind speed and vibration intensity are extracted from sensor data. To ensure time synchronization across different data sources, a timestamp alignment method is used to map all feature data onto a unified time grid, with a time granularity of 1 minute. In terms of feature engineering, this embodiment designs multi-scale spatiotemporal features: in the spatial dimension, not only are local features of individual grid cells extracted, but features of neighboring nodes are also aggregated through graph convolution; in the temporal dimension, short-term (minute-level) and long-term (hour-level) dependencies are captured through a multi-head self-attention mechanism.

[0021] Step 3: Construct a spatiotemporal graph convolutional network model (model structure as follows) Figure 2 (As shown). This model consists of alternating stacks of spatial graph convolutional modules and temporal convolutional modules. The spatial graph convolutional module uses graph convolution operations to aggregate node features, and the calculation formula is: ,in This represents the feature matrix of the nodes in the l-th layer. To add a self-loop adjacency matrix, for The degree matrix, The weight matrix is ​​learnable, and σ is the activation function. The temporal convolution module uses one-dimensional convolution operations to capture temporal dependencies.

[0022] Step 4: Model Training and Optimization. Supervised learning training is performed on the model using historical labeled data. The loss function is a combination of cross-entropy loss and L2 regularization. An early stopping strategy is employed during training to prevent overfitting: training stops when the validation set loss does not decrease for 10 consecutive epochs. After model training is complete, predictive performance is evaluated on the test set.

[0023] Step 5: Real-time Risk Prediction and Early Warning. The multi-source data collected in real time is input into the trained ST-GCN model. The model outputs the fall risk probability for each grid node in future time periods (e.g., 15 minutes, 30 minutes, 1 hour). When the predicted probability exceeds a preset threshold (e.g., a high-risk threshold of 0.8), the system automatically generates an early warning message and visualizes the risk distribution as a heatmap on the digital twin model. This embodiment also uses Grad-CAM technology to interpret the model's prediction results to identify the key factors that contribute most to the prediction results.

[0024] Step 6: Early Warning Information Push and Task Distribution (Related procedures can be found in the original text) Figure 5 and Figure 6 Based on the warning level and the scope of impact, the system automatically matches emergency response plans and distributes response tasks to relevant responsible persons.

[0025] Data robustness enhancement mechanism: In actual construction sites, the data acquisition environment is often less than ideal. Sensors may experience data loss or anomalies due to vibration, dust, unstable power supply, etc., and communication links may be frequently interrupted due to the complex electromagnetic environment of the construction site. To ensure the reliability of the system under real-world conditions, this system is designed with the following data quality assurance mechanism.

[0026] Specifically, the system performs real-time quality scoring on each data source. The scoring dimensions include data freshness (timestamp delay), completeness (missing rate), stability (variance mutation detection), and physical plausibility (value range and rate of change threshold). The quality score is dynamically updated using a weighted moving average method, with a score range of 0 to 1. The quality score calculation formula is as follows: ,in The default weight coefficients for each dimension are 0.3, 0.2, 0.3, and 0.2, respectively, and can be dynamically adjusted according to the specific sensor type. During the multi-source data fusion process, the system automatically adjusts the fusion weights based on the quality scores of each data source: the weight of low-quality data is reduced, and in extreme cases (scores below 0.2), the data source is temporarily blocked and an operation and maintenance alarm is triggered.

[0027] For short-term data gaps (e.g., fewer than 10 consecutively missing sampling points), the system uses a spatiotemporal K-nearest neighbor imputation method: finding adjacent similar nodes (K=5) on the spatiotemporal graph and using their observations to infer the missing value through inverse distance weighting. For long-term gaps (≥10 consecutively missing sampling points) or multi-dimensional joint gaps, a pre-trained Conditional Generative Adversarial Network (cGAN) is used to generate complete data. This module uses the missing mask and spatiotemporal neighborhood observations as conditions, and a discriminator ensures that the completed data is consistent with the distribution of the real data, with a single completion taking less than 50 milliseconds.

[0028] Outlier detection employs a cascaded approach based on Isolation Forest and the 3σ criterion: First, a sliding window (window size = 100 samples) is used to calculate the mean and standard deviation for each data stream. Points deviating from the mean by more than three times the standard deviation are marked as candidate outliers. Then, the candidate outliers are further confirmed using an Isolation Forest model, which is incrementally trained every 24 hours on the edge server. Points confirmed as outliers are not discarded directly but are corrected based on data type: for slowly changing data (such as temperature and stress), median filtering is used; for rapidly changing data (such as vibration and current), linear interpolation or replacement with nearby normal values ​​is used. All correction operations are logged.

[0029] In addition, the system deploys dual-mode communication (4G / 5G + LoRa) in key areas on-site. Under normal circumstances, high-bandwidth 5G is used to transmit image and point cloud data, and when the 5G signal is interrupted, it automatically switches to LoRa to transmit structured feature data. The edge server is configured with 64GB of local cache, which can store 72 consecutive hours of raw and feature data. When cloud communication is completely interrupted, the edge server independently runs a lightweight risk prediction model. This lightweight model reduces the number of parameters to 30% of the original model by maintaining the overall architecture of ST-GCN, reducing the number of hidden units in each graph convolution layer from 64 to 32, the number of temporal convolution kernels from 128 to 64, and removing the last two fully connected layers, replacing them with a single-layer linear mapping output. This lightweight model achieves an F1 score of 92% of the original model on the test set, with an inference speed improvement of 2.3 times, which is sufficient to ensure basic early warning capabilities during network outages. After communication is restored, the system automatically synchronizes all data from the network outage period and triggers incremental model updates. This mechanism ensures uninterrupted on-site safety management under poor network conditions. Field tests show that even with 30% of the sensors malfunctioning, the system can still maintain a prediction accuracy of over 85%.

[0030] Data flow direction (for reference) Figure 3 (Multi-source data fusion flowchart): After raw data is collected from various sensors and cameras, it is first transmitted to an edge computing server for preprocessing and feature extraction. The extracted feature data is then transmitted to a cloud server via a message queue. The cloud server runs the ST-GCN model for risk prediction. The prediction results output by the model are written to a time-series database, simultaneously triggering an early warning service. The early warning service generates visual information and task push instructions based on the prediction results, which are ultimately pushed to relevant users via web and mobile devices. The entire data flow process uses Kafka message queues and Redis caching technology to ensure the real-time performance and reliability of data transmission.

[0031] Example 2: Multi-level Early Warning and Intelligent Decision Support System Based on Digital Twin (Metro Construction) This embodiment takes the safety management of a large-scale subway construction project as an example to illustrate the specific application of a multi-level early warning and intelligent decision support system based on digital twins. Figure 4 This is a schematic diagram illustrating the mapping between a digital twin model and real-time data. Figure 5 A flowchart for multi-level early warning and intelligent decision support. Figure 6 This diagram illustrates a coordinated emergency response and closed-loop management mechanism. Subway construction involves complex processes such as deep foundation pit excavation, tunnel boring, and underground structure construction. Safety risks include various types such as collapse, water inrush, and gas poisoning, and these risks may influence and transmit each other.

[0032] In terms of digital twin modeling, a BIM model based on the IFC (Industry Foundation Classes) standard was adopted as the geometric and semantic foundation. The model contains complete information such as soil structure, support system, and monitoring point layout. Instantiation technology and Level of Detail (LOD) optimization technology were used to control the number of polygon patches in the model to within 5 million, and the rendering frame rate reached more than 60 frames per second. In terms of data integration, an automated monitoring robot (AMR) was deployed with a monitoring accuracy of 0.1 mm and a monitoring frequency of once every 2 hours; pore water pressure gauges, earth pressure cells, and other sensors were deployed inside the foundation pit with a sampling frequency of 1 Hz; methane sensors and carbon monoxide sensors were deployed at the tunnel excavation face for real-time monitoring of harmful gas concentrations.

[0033] Implementation steps: Step 1: Digital Twin Model Construction and Lightweighting. Based on design drawings and BIM software, a 3D geological model and structural model were created. The model was converted to glTF format, and geometric compression was performed using the Draco compression algorithm. Instantiation technology was used to reduce the storage overhead of repetitive geometries. After lightweighting, the model file size was compressed from the original 5GB to less than 500MB.

[0034] Step 2: Multi-source data integration and real-time mapping (e.g.) Figure 4 (As shown). A semantic association is established between the BIM model and monitoring points, ensuring that each monitoring point has a corresponding geometric object in the model. When monitoring data arrives, the system automatically updates the status attributes of the corresponding objects in the model according to predefined mapping relationships, for example, visually displaying abnormal states through color changes.

[0035] Step 3: Safety Situation Assessment and Tiered Early Warning. The system employs a multi-factor weighted fusion method to comprehensively assess the on-site safety situation. Assessment indicators include the degree of exceeding monitoring limits, rate of change, duration, and surrounding environmental factors. Based on the assessment results, the safety status is divided into four levels: Level 1 (Normal, Green), Level 2 (Attention, Yellow), Level 3 (Warning, Orange), and Level 4 (Emergency, Red). Different levels correspond to different response measures. For example, Level 2 requires increased monitoring frequency, while Level 3 requires suspension of relevant operations and activation of emergency plans.

[0036] Step 4: Intelligent Decision Support and Solution Recommendation. After the system issues an alert, the decision support module retrieves similar historical cases from the knowledge base based on factors such as the current risk type, severity, and on-site conditions, and generates an emergency response plan based on rule-based reasoning.

[0037] Detailed implementation of the knowledge base: The safety knowledge base of this invention adopts a hybrid storage architecture combining graph database (Neo4j) and vector database (Milvus) to support efficient case reasoning and semantic similarity retrieval. The knowledge base is built on a safety domain ontology based on OWL (Web Ontology Language), with core concepts including: risk events (falls, collapses, electric shocks, mechanical injuries, object strikes, etc.), risk factors (environmental factors, behavioral factors, equipment factors), response measures (evacuation, reinforcement, work stoppage, enhanced monitoring, activation of emergency plans), and resources (emergency personnel, equipment and supplies, evacuation routes). The ontology contains approximately 120 classes and 350 attribute constraints, supporting descriptive logic (DL) reasoning.

[0038] Each safety case is represented as a triple <scenario feature vector, handling plan, effect feedback>. The scenario feature vector is extracted from historical early warning data and contains 48 features across 6 categories: risk type encoding (one-hot), spatiotemporal graph features (graph structure hashing), monitoring data statistics (mean / variance / trend), environmental parameters, construction stage, and personnel density. Cases employ a dual index: the graph index stores the relationships between cases in Neo4j; the vector index uses a pre-trained Siamese network to map the 48-dimensional features to a 128-dimensional embedding vector, creating an IVF_FLAT index in Milvus, supporting millisecond-level cosine similarity retrieval.

[0039] The hybrid inference mechanism executes the following process: First, the system attempts to match predefined deterministic rules. The rule base contains, for example, 2150 rules (this number grows dynamically with project accumulation; initially, approximately 800 core rules were written by security experts, and it will be expanded through continuous learning), covering common scenarios. The rules employ a forward-chain inference engine (based on the Rete algorithm), for example: "IF Risk Type = Excavation Pit Deformation AND Deformation Rate > 5 mm / day AND Cumulative Deformation > 0.8%H" THEN recommends: Immediately stop excavation, install additional steel supports, and increase monitoring frequency to once per hour.

[0040] If a match is successful and the rule confidence score is >0.9, the solution is output directly. If the rule match fails or the confidence score is insufficient, the system switches to case-based reasoning (CBR), retrieving the Top-K historical cases (K=5) in Milvus that are most similar to the current scene's feature vector, using weighted cosine similarity: Where Δt represents the time since the case (in days), λ is set to 0.01, and the bonus factor for similar projects is 0.1. Similar cases retrieved may have different handling measures; the system uses majority voting combined with conflict resolution rules to generate recommended solutions (e.g., the principle of "safety over schedule"). The final output solution includes: a list of handling actions, responsible roles, required resources, and expected completion time.

[0041] The knowledge base employs a continuous learning mechanism: After each security incident is handled, the system automatically packages the scenario characteristics of the alert, the actual handling plan adopted, and the handling effect (based on subsequent monitoring data and manual evaluation) into a new case and inserts it into Neo4j and Milvus. If the similarity between a new case and an existing case is >0.95, and the effect score is higher, the old case is marked as "outdated," and its weight is reduced during inference. A weekly association rule mining (Apriori algorithm) is run to extract high-frequency feature combinations and their associations with handling measures from successfully handled cases, generating candidate new rules, which are then added to the rule base after review by security experts. The Siamese network is fine-tuned quarterly using the updated case base. Through this mechanism, the knowledge base's ability to cover newly added scenarios with rules increases by approximately 5% monthly, and the case retrieval accuracy remains stable above 92%.

[0042] Step 5: 3D Visualization and Interactive Operation. The digital twin platform provides functions such as profile viewing, contour line generation, and deformation trend animation, and supports spatiotemporal backtracking: users can drag the time slider to view the on-site status at any historical moment.

[0043] Four-dimensional spatiotemporal modeling: This embodiment introduces a four-dimensional spatiotemporal modeling method into the construction of the digital twin model, that is, adding a time dimension in addition to the geometric dimension. By associating the BIM model with the construction schedule (4D), dynamic simulation and deduction of the construction process can be achieved. The system supports multi-view linkage display and linkage display of two-dimensional charts and three-dimensional models.

[0044] Model comparison and validation and ablation experiments To verify the superiority of the proposed ST-GCN model in construction safety risk prediction, this embodiment collected six months of data (including 127 labeled samples of three types of risk events: falls, object strikes, and collapses) in an actual construction site deployment environment, and constructed a standard test set. It was compared with the following four representative methods: Baseline 1 is a single-threshold alarm model (a typical representative of existing technologies); Baseline 2 is an LSTM temporal prediction model (utilizing only the time dimension, without spatial correlation); Baseline 3 is a GCN graph convolution model (utilizing only the spatial graph structure, without temporal convolution); Baseline 4 is a ConvLSTM model (using convolutional LSTM on a regular grid, without using a graph structure). Experimental settings: the training set / validation set / test set were divided into 70% / 15% / 15% ratios. Evaluation metrics included accuracy, recall, F1 score, false alarm rate (FAR), and average warning lead time (minutes). All models used the same multi-source input features, and the output was the risk probability for the next 30 minutes (a threshold of 0.7 was considered a warning).

[0045] The experimental results are shown in the table below: To further verify the contributions of each key module in this invention, five ablation variants were designed: full ST-GCN (spatiotemporal graph convolution + dynamic graph + multi-scale features + data robustness); without dynamic graph (using a static adjacency matrix); without multi-scale features (using only single-scale spatiotemporal features); without data robustness (excluding mechanisms such as missing data handling and anomaly correction); without knowledge augmentation (not using knowledge base decision support, relying solely on model probability output). Ablation experiments show that: dynamic graph updates improve the F1 score from 0.78 to 0.85; multi-scale features improve the F1 score from 0.80 to 0.85; the data robustness module improves the F1 score from 0.72 to 0.85 in noisy environments; and knowledge augmentation significantly improves decision accuracy. The proposed ST-GCN model significantly outperforms existing methods in all metrics, especially in early warning time for high-risk events, improving it by approximately 7.7 minutes compared to the optimal baseline (ConvLSTM).

[0046] Through the above architecture and methods, a closed loop of comprehensive perception, intelligent analysis, and collaborative control for construction safety management has been achieved.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0048] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0049] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-source heterogeneous data fusion-based integrated digital and intelligent collaborative system for group construction enterprises, characterized in that: include: The data acquisition layer is used to acquire video surveillance data, IoT sensor data, personnel positioning data, equipment status data, and environmental monitoring data from the construction site. The data processing layer is used to perform spatiotemporal alignment, cleaning, fusion, and feature extraction on the multi-source heterogeneous data. The model computation layer contains a risk prediction model based on a spatiotemporal graph convolutional network (ST-GCN), which is used for spatiotemporal correlation analysis and probability prediction of construction safety risks. The application service layer is used to trigger multi-level early warnings, generate emergency response plans, and intelligently distribute tasks based on the prediction results; The interactive layer displays a 3D visualization interface and human-computer interaction functions based on digital twin technology.

2. The system according to claim 1, characterized in that, The spatiotemporal graph convolutional network adopts a dynamic graph update mechanism: based on the personnel location and equipment status at the construction site, the adjacency relationship of the spatiotemporal graph is dynamically updated at preset intervals; the spatiotemporal graph nodes correspond to the grid cells divided at the construction site, and the edges represent spatial adjacency relationships or functional association relationships; the network structure includes alternating stacked graph convolutional layers and temporal convolutional layers, wherein the graph convolutional layers use a normalized adjacency matrix to aggregate node features, and the temporal convolutional layers use one-dimensional convolution to capture temporal dependencies.

3. The system according to claim 1, characterized in that, The data processing layer includes a data robustness enhancement module, which performs the following operations: A quality score is calculated in real time for each data source, with scoring dimensions including data freshness, completeness, stability, and physical reasonableness. The data fusion weights are dynamically adjusted based on the scores. For short-term missing data, a spatiotemporal K-nearest neighbor imputation method is used; for long-term missing or multi-dimensional joint missing data, data completion based on conditional generative adversarial networks (cGAN) is used. Outliers are detected by cascading isolation forest and 3σ criterion, and median filtering or linear interpolation correction is performed according to data type. Configure local caching on the edge server so that the lightweight risk prediction model can run independently on the edge server when cloud communication is interrupted.

4. The system according to claim 1, characterized in that, The interactive display layer constructs a digital twin model based on BIM, GIS, and on-site 3D scanning data. It uses instantiation technology and Level of Detail (LOD) optimization to achieve model lightweighting, maps monitoring data to model objects in real time, and visualizes the distribution of safety risks through color changes, heat maps, animations, and other methods. It supports four-dimensional spatiotemporal modeling and spatiotemporal backtracking functions.

5. The system according to claim 1, characterized in that, The application service layer adopts a multi-level early warning mechanism: based on the probability of risk, the degree of impact and the urgency, the security status is divided into four levels: normal, attention, early warning and emergency, with each level corresponding to different response measures; After an early warning is triggered, the system provides intelligent decision support based on a knowledge base. The knowledge base uses a hybrid storage method combining graph databases and vector databases, and includes deterministic rules and typical cases, supporting a hybrid reasoning mechanism that combines rule-based reasoning with case-based reasoning.

6. The system according to claim 5, characterized in that, The knowledge base employs a continuous learning mechanism: after each security incident is handled, the system automatically adds the scene features, handling plan, and effect feedback as new cases to the database; it periodically runs an association rule mining algorithm to extract new rules from successful cases; it cleans up redundant outdated cases and periodically fine-tunes the feature embedding model.

7. The system according to claim 1, characterized in that, The application service layer also includes an emergency task intelligent distribution module based on roles and locations: it calculates the matching degree between tasks and responsible persons according to the warning level, task type and on-site personnel location information, automatically generates task push strategies, tracks task execution status in real time, and automatically upgrades overdue tasks.

8. The system according to claim 1, characterized in that, The system adopts a cloud-edge collaborative architecture: the edge is responsible for data preprocessing and lightweight model inference, while the cloud is responsible for complex model training and global data analysis; the model training adopts a transfer learning strategy, first training a general model on historical project data, and then fine-tuning it on specific projects using a small amount of labeled data.

9. A method for predicting construction safety risks of the system according to any one of claims 1 to 8, characterized in that, Includes the following steps: Construct a spatiotemporal graph structure for the construction site, dividing the construction area into multiple grid units as graph nodes; Collect multi-source heterogeneous data and perform spatiotemporal alignment and feature extraction; The extracted features are input into the trained spatiotemporal graph convolutional network model, which outputs the probability of each node falling in the future. When the predicted probability exceeds a preset threshold, an early warning message is generated, and the risk distribution is visualized on the digital twin model. Based on the warning level and the scope of impact, emergency response plans are automatically matched and distributed to relevant responsible persons.

10. The method according to claim 9, characterized in that, During the training process of the spatiotemporal graph convolutional network, the loss function adopts a combination of cross-entropy loss and L2 regularization term, and an early stopping strategy is used to prevent overfitting. The model prediction results are analyzed for interpretability using Grad-CAM technology to identify the key factors that contribute the most to the prediction results.

Images